Re: [PATCH 2/3] asm-generic, x86: wrap atomic operations

From: Dmitry Vyukov <dvyukov@google.com>
To: Mark Rutland <mark.rutland@arm.com>
Cc: Andrey Ryabinin <aryabinin@virtuozzo.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>, Will Deacon <will.deacon@arm.com>,
	Andrew Morton <akpm@linux-foundation.org>,
	kasan-dev <kasan-dev@googlegroups.com>,
	"linux-mm@kvack.org" <linux-mm@kvack.org>,
	"x86@kernel.org" <x86@kernel.org>,
	LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 2/3] asm-generic, x86: wrap atomic operations
Date: Tue, 21 Mar 2017 19:06:14 +0100	[thread overview]
Message-ID: <CACT4Y+bNrh_a8mBth7ewHS-Fk=wgCky4=Uc89ePeuh5jrLvCQg@mail.gmail.com> (raw)
In-Reply-To: <20170321104139.GA22188@leverpostej>

On Tue, Mar 21, 2017 at 11:41 AM, Mark Rutland <mark.rutland@arm.com> wrote:
> On Tue, Mar 21, 2017 at 12:25:06PM +0300, Andrey Ryabinin wrote:
>> On 03/20/2017 08:17 PM, Mark Rutland wrote:
>> > Hi,
>> >
>> > On Tue, Mar 14, 2017 at 08:24:13PM +0100, Dmitry Vyukov wrote:
>> >>  /**
>> >> - * atomic_read - read atomic variable
>> >> + * arch_atomic_read - read atomic variable
>> >>   * @v: pointer of type atomic_t
>> >>   *
>> >>   * Atomically reads the value of @v.
>> >>   */
>> >> -static __always_inline int atomic_read(const atomic_t *v)
>> >> +static __always_inline int arch_atomic_read(const atomic_t *v)
>> >>  {
>> >> -  return READ_ONCE((v)->counter);
>> >> +  /*
>> >> +   * We use READ_ONCE_NOCHECK() because atomic_read() contains KASAN
>> >> +   * instrumentation. Double instrumentation is unnecessary.
>> >> +   */
>> >> +  return READ_ONCE_NOCHECK((v)->counter);
>> >>  }
>> >
>> > Just to check, we do this to avoid duplicate reports, right?
>> >
>> > If so, double instrumentation isn't solely "unnecessary"; it has a
>> > functional difference, and we should explicitly describe that in the
>> > comment.
>> >
>> > ... or are duplicate reports supressed somehow?
>>
>> They are not suppressed yet. But I think we should just switch kasan
>> to single shot mode, i.e. report only the first error. Single bug
>> quite often has multiple invalid memory accesses causing storm in
>> dmesg. Also write OOB might corrupt metadata so the next report will
>> print bogus alloc/free stacktraces.
>> In most cases we need to look only at the first report, so reporting
>> anything after the first is just counterproductive.
>
> FWIW, that sounds sane to me.
>
> Given that, I agree with your comment regarding READ_ONCE{,_NOCHECK}().
>
> If anyone really wants all the reports, we could have a boot-time option
> to do that.

I don't mind changing READ_ONCE_NOCHECK to READ_ONCE. But I don't have
strong preference either way.

We could do:
#define arch_atomic_read_is_already_instrumented 1
and then skip instrumentation in asm-generic if it's defined. But I
don't think it's worth it.

There is no functional difference, it's only an optimization (now
somewhat questionable). As Andrey said, one can get a splash of
reports anyway, and it's the first one that is important. We use KASAN
with panic_on_warn=1 so we don't even see the rest.